Date of Graduation


Document Type


Degree Type



Statler College of Engineering and Mineral Resources


Lane Department of Computer Science and Electrical Engineering

Committee Chair

Tim Menzies.


There exists a large number of software effort estimation methods in the literature and the space of possibilities [54] is yet to be fully explored. There is little conclusive evidence about the relative performance of such methods and many studies suffer from instability in their conclusions. As a result, the effort estimation literature lacks a stable ranking of such methods.;This research aims at providing a stable ranking of a large number of methods using data sets based on COCOMO features. For this task, the COSEEKMO tool [46] was further developed into a benchmarking tool and several well-known effort estimation methods, including model trees, linear regression methods, local calibration, and several newly developed methods were used in COSEEKMO for a thorough comparison. The problem of instability was further explored and the evaluation method used was identified as the cause of instability. Therefore, the existing evaluation bias was corrected through a new evaluation approach, which was non-parametric. The Mann-Whitney U test [42] is the non-parametric test used in this study, which introduced a great amount of stability in the results. Several evaluation criteria were tested in order to analyze their possible effects on the observed stability.;The conclusions made in this study were stable across different evaluation criteria, different data sets, and different random runs. As a result, a group of four methods were selected as the best effort estimation methods among the explored 312 combinations of methods. These four methods were all based on the local calibration procedure proposed by Boehm [4]. Furthermore, these methods were simpler and more effective than many other complex methods including the Wrapper [37] and model trees [60], which are well-known methods in the literature.;Therefore, while there exists no single universal best method for effort estimation, this study suggests applying the four methods reported here to the historical data and using the best performing method among these four to estimate the effort for future projects. In addition, this study provides a path for comparing other existing or new effort estimation methods with the currently explored methods. This path involves a systematic comparison of the performance of each method against all other methods, including the methods studied in this work, through a benchmarking tool such as COSEEKMO, and using the non-parametric Mann-Whitney U test.